Gas turbine remaining useful life prediction method based on improved SFOA algorithm

Abstract

The application provides a gas turbine residual service life prediction method based on an improved SFOA algorithm, and belongs to the technical field of gas turbine life prediction, in particular: collecting gas turbine life cycle operation parameters and forming a data set; combining a space-time convolution network, a bidirectional long short-term memory neural network and a multi-head attention mechanism to form an initial prediction model; using a magnificent fine-tailed wren optimization algorithm to optimize model parameters of the initial prediction model to obtain an optimized prediction model, training the optimized prediction model, collecting real-time operation parameters of a current gas turbine to be predicted, extracting features of the operation parameters, processing the extracted features, inputting the processed features into the trained prediction model, and outputting a gas turbine residual service life prediction result by the trained prediction model. Through the processing scheme, the prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of gas turbine life prediction, and in particular to a gas turbine remaining useful life prediction method based on an improved SFOA algorithm. BACKGROUND

[0002] As the core power device of gas power generation, aviation and ship, the gas turbine occupies a key position in transportation and livelihood security. Its design, manufacture, operation management and maintenance capability effectively reflects the scientific and technological development level of a country. Due to the complex structure of the gas turbine, and often running in harsh environments such as high temperature and high pressure, combined with frequent start and stop, it is easy to cause the failure of each key component in the gas turbine, and then seriously affect its service life. By predicting the remaining useful life of the gas turbine, the maintenance decision can be provided, and the maintenance plan can be made in advance, so as to effectively prolong the equipment life, reduce the unplanned downtime, and reduce the operation and maintenance cost.

[0003] At present, there are mainly two technical paths for gas turbine remaining life prediction: a prediction method based on a physical model and a prediction method based on data driving. The physical model method mainly analyzes the structure and working principle of the gas turbine, and establishes a degradation mechanism model to realize life prediction. The data driving method constructs a mapping model by using historical working condition data, and realizes the prediction function by various intelligent algorithms. In recent years, with the increasing complexity of the structure of the gas turbine, it becomes more and more difficult to construct an accurate physical model when using the physical model method, and the data driving method thus shows a significant advantage. In terms of improving the prediction accuracy, the traditional machine learning algorithm has certain limitations, and deep learning has strong generalization performance because it can automatically extract the deep attributes and hidden features of data, and has the advantages of not relying on mechanism modeling, etc. It has become the mainstream method of gas turbine life prediction. However, the traditional deep learning prediction method has the problems of difficult selection of hyperparameters and insufficient prediction accuracy. SUMMARY

[0004] Therefore, the present application provides a gas turbine remaining useful life prediction method based on an improved SFOA algorithm, which solves the problems in the prior art and improves the prediction accuracy.

[0005] The gas turbine remaining useful life prediction method based on the improved SFOA algorithm provided by the present application adopts the following technical scheme:

[0006] A gas turbine remaining useful life prediction method based on an improved SFOA algorithm, comprising:

[0007] Step 1, collect the operating parameters of the gas turbine throughout its life cycle, extract the characteristics of the operating parameters, process the extracted characteristics, and label the processed characteristics with corresponding remaining useful life labels to form a data set;

[0008] Step 2, build an initial prediction model, the initial prediction model includes a space-time convolution network, a bidirectional long short-term memory neural network, and a multi-head attention mechanism, the space-time convolution network as a front-end feature extractor, the bidirectional long short-term memory neural network as a time series dependence modeling module, and the multi-head attention mechanism as a global information integration module;

[0009] Step 3, use the magnificent fine-tailed warbler optimization algorithm to optimize the model parameters of the initial prediction model, take the learning rate, the number of space-time convolution network neurons, the number of bidirectional long short-term memory neural network neurons, the number of multi-head attention mechanism keys, and the regularization parameter as optimization variables, and take the root mean square error of the prediction results of the initial prediction model on the data set as the optimization objective function, and minimize the root mean square error to obtain an optimized prediction model;

[0010] Step 4, divide the data set into a training set and a test set, train the optimized prediction model until the model converges, and obtain a trained prediction model;

[0011] Step 5, collect real-time operating parameters of the current gas turbine to be predicted, extract the characteristics of the operating parameters, process the extracted characteristics, input the processed characteristics into the trained prediction model, and output the remaining useful life prediction result of the gas turbine.

[0012] Optionally, the step of running the trained prediction model in step 5 includes:

[0013] Step 5.1, the space-time convolution network uses a causal convolution layer, an expansion convolution layer, and a residual connection to extract memory history features of time series data;

[0014] Step 5.2, the bidirectional long short-term memory neural network receives the memory history features of the time series data, simultaneously processes the forward information and the reverse information of the memory history features of the time series data, captures the time global degradation dependence relationship, and fuses the features to output global information;

[0015] Step 5.3, the multi-head attention mechanism receives the global information, assigns different importance weights through multiple independent attention heads, excavates the degradation time series information in the global information, and outputs the remaining useful life prediction result of the gas turbine.

[0016] Optionally, the space-time convolution network includes a plurality of residual blocks, each residual block includes a causal convolution layer, an expansion convolution layer, and a residual connection;

[0017] The causal convolution layer is used for performing a causal convolution operation on a time sequence vector, a convolution kernel size of the causal convolution layer is 3*3, and a step length is 1;

[0018] The dilated convolution layer is used for performing a dilated convolution operation on an output result of the causal convolution layer, a dilated coefficient of the dilated convolution layer is 2 a-1 is set, wherein a is a convolution layer serial number, a = 1, 2, 3, 4, 5, so that the receptive field is exponentially expanded with the increase of the layer number;

[0019] The residual connection fuses the output of each dilated convolution layer and the input of the residual block through a shortcut connection mode, and integrates the input features and the convolution output features in an element addition mode; if the dimensions of the input and the output are inconsistent, the dimensions are adjusted through 1*1 convolution before the addition operation is performed.

[0020] Optionally, the bidirectional long short-term memory neural network in step 5.2 refines the memory history features of the time sequence data by sharing the sequence of the memory history features of the output time sequence data and processing in the opposite time sequence, and finally outputs global information by fusing the features;

[0021] The bidirectional long short-term memory neural network includes three layers of bidirectional LSTM units, and the number of neurons of the forward LSTM and the backward LSTM in each layer is the same;

[0022] The forward LSTM traverses the input sequence in the positive direction of the time sequence, updates the cell state and the hidden state through the gating unit, generates a positive hidden state sequence, and captures the positive time dependence of the sequence;

[0023] The backward LSTM traverses the same input sequence in the reverse direction of the time sequence, updates the cell state and the hidden state through the gating unit, generates a reverse hidden state sequence, and captures the reverse time dependence of the sequence.

[0024] Optionally, the step 5.3 specifically includes:

[0025] Target input receiving: the multi-head attention mechanism receives the feature fusion output global information as the target input of the multi-head attention mechanism;

[0026] Feature encoding: if it is the first layer processing of the multi-head attention mechanism, linear encoding is performed on the target input to generate a feature matrix H, and if it is non-first layer processing, the output of the last layer of multi-head attention is directly taken as the feature matrix H;

[0027] Multi-head weighting: Construct h independent linear projection matrices, h≥8. The linear projection matrices include query projection matrix group Wq, key projection matrix group Wk, and value projection matrix group Wv. Multiply the feature matrix H with each group Wq, Wk, and Wv respectively to obtain h query matrices Q1-Q. h H group bond matrix K1-K h h group value matrix V1-V h ;

[0028] Single-head attention calculation: For each combination matrix (Q) m ,K m V m The attention weights are calculated using the scaled dot product attention formula, resulting in h groups of single-head attention outputs Z1-Z. h , where m takes the value of an integer from 1 to h;

[0029] Multi-head result stitching and projection: The h groups of single-head attention outputs Z1-Z h Concatenate the columns to obtain the concatenation matrix Z_concat; perform linear projection on the concatenation matrix Z_concat, multiply the concatenation matrix Z_concat by the learned weight matrix, reduce the dimension to the preset output dimension, and obtain the output matrix Z_final;

[0030] Prediction results output: Based on the projected output matrix Z_final, extract global key degradation time series information, and combine it with the preset remaining life mapping rules to output the prediction results of the remaining life of the gas turbine.

[0031] Optionally, the optimization algorithm steps for the magnificent swan-warbler in step 3 include:

[0032] Step 31: Convert the natural morphology of the magnificent swan warbler in nature into a mathematical model. Each individual magnificent swan warbler corresponds to a set of model parameters to be optimized in the initial prediction model. Each dimension corresponds to an optimization variable. Use a sinusoidal squared chaotic mapping to generate an initial candidate solution population in the search space. }, ;

[0033] The formula for the sinusoidal squared chaotic mapping is: ;

[0034] in, For the first The chaotic variable after the next iteration is initially set to a random decimal between 0 and 1. For the first The updated value after the next iteration; For control parameters; For the initial candidate solution population, the first The first individual Parameter values ​​for each dimension; This is the upper bound of the search space; This serves as the lower bound of the search space.

[0035] Step 32: Determine the proportion of juveniles in the initial candidate solution population. If the proportion of juveniles is greater than 0.5, update the individual parameter values ​​in the initial candidate solution population by simulating the growth stage of juvenile magnate warbler until the proportion of juveniles in the updated initial candidate solution population is less than or equal to 0.5, and obtain a qualified candidate solution population.

[0036] If the proportion of juveniles in the initial candidate solution population is less than or equal to 0.5, then the initial candidate solution population is directly regarded as a qualified candidate solution population.

[0037] Step 33: Calculate the risk threshold s in the qualified candidate solution population. If the risk threshold is less than a preset value, proceed to step 34; if the risk threshold is greater than or equal to the preset value, proceed to step 35. , and It is a random number that follows a normal distribution;

[0038] Step 34: Select the optimal solution individual from the qualified candidate solution population by simulating the teaching mechanism of the magnificent slender-tailed warbler in the process of breeding and feeding. The model parameter combination corresponding to the optimal solution individual is used as the model parameters of the optimized prediction model.

[0039] Step 35: Select the optimal solution individual from the qualified candidate solution population by simulating the mechanism of the magnificent slender-tailed warbler in avoiding predators. The model parameter combination corresponding to the optimal solution individual is used as the model parameters of the optimized prediction model.

[0040] Optionally, step 35 specifically includes:

[0041] Taking the individual with the best optimization objective function in the qualified candidate solution population as the subject, the parameters of the remaining individuals in the qualified candidate solution population are adjusted according to the position update rules of the magnificent swan warbler in the process of simulating the mechanism of avoiding natural enemies, and the parameter value combination of each individual is updated.

[0042] The updated formula is: ;

[0043] in, The first qualified candidate solution in the population The first individual The updated parameter values ​​for each dimension. This refers to the optimal combination of individual parameter values ​​for optimizing the objective function within the qualified candidate solution population. express Individualst After the second iteration j Parameter values ​​for each dimension;

[0044] The random step size during flight. , for Random values ​​in;

[0045] For adaptive flight balance factor, , , For call frequency values, Indicates the current evaluation number. Indicates the maximum number of evaluations;

[0046] If the objective function corresponding to the updated parameter value combination of the remaining individuals in the qualified candidate solution population is better than the objective function corresponding to the original parameter value combination, then the updated parameter value combination of the individual replaces the original parameter value combination of the individual. After all the remaining individuals complete the preset number of updates, a new qualified candidate solution population is obtained. From the new qualified candidate solution population, the individual with the optimal objective function is selected, and the model parameter combination corresponding to the optimal individual is used as the model parameters of the final optimized prediction model.

[0047] Optionally, in steps 1 and 5, the step of extracting features of the operating parameters includes:

[0048] Extract the time-domain and frequency-domain features of the operating parameters. The time-domain features include maximum value, minimum value, average value, peak value, peak-to-peak value, rectified average value, variance, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, root mean square amplitude, margin factor, and kurtosis factor.

[0049] Frequency domain characteristics include centroid frequency, mean square frequency, frequency variance, spectral entropy, and signal complexity.

[0050] Optionally, in steps 1 and 5, the step of processing the extracted features includes performing sliding window sampling on the extracted features;

[0051] The sliding window spans 5-30 sampling periods, and the window sliding step is 1 sampling period.

[0052] The features extracted from real-time operating parameters are completely consistent with the feature types extracted from the whole life cycle data, ensuring that the input distribution matches.

[0053] In summary, this application includes the following beneficial technical effects:

[0054] This application combines a temporal convolutional network (TCNN), a bidirectional long short-term memory (LSTM) neural network, and a multi-head attention mechanism—all optimized by the Magnificent Wren's algorithm—to predict the remaining service life of gas turbines, thereby enabling fault identification. This application achieves perfect complementarity between the multi-head attention mechanism and the TCNN, constructing a classic model that leverages the strengths of both TCNN and LSTM, overcoming the shortcomings of the original TCNN and LSTM models. This allows the entire model to perform global modeling in one step, fully utilizing the massive parallel computing capabilities of GPUs and other hardware during training, significantly accelerating the training speed, completely resolving the inherent bottlenecks of the TCNN series of models, and ultimately making accurate predictions.

[0055] This application also employs an improved Slender-tailed Warbler optimization algorithm, which enhances the model's global exploration capability in the early stages and strengthens its local exploration capability in the later stages, while reducing the model's sensitivity to initial parameters. The Slender-tailed Warbler optimization algorithm enables the model to find a better solution in a shorter time, avoiding getting trapped in local optima. The ultimate effect of this invention is to improve the accuracy of gas turbine life prediction, solve the inherent problems of data-driven prediction methods and traditional temporal convolutional network models, and address the insufficient prediction accuracy of other life prediction methods. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 The flowchart shows a method for predicting the remaining useful life of a gas turbine based on an improved SFOA algorithm.

[0058] Figure 2 This is a structural block diagram of the spatiotemporal convolutional network in this application;

[0059] Figure 3 This is a block diagram of a bidirectional long short-term memory neural network structure. Detailed Implementation

[0060] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0061] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0063] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The illustrations only show the components related to this application and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0064] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0065] This application provides a method for predicting the remaining service life of a gas turbine based on an improved SFOA algorithm.

[0066] like Figure 1 As shown, a method for predicting the remaining service life of a gas turbine based on an improved SFOA algorithm includes:

[0067] Step 1, Constructing a dataset: Collect operating parameters of the gas turbine throughout its entire life cycle, extract features from the operating parameters, process the extracted features, and label the processed features with the corresponding remaining service life labels to form a dataset.

[0068] Step 2: Construct an initial prediction model, which includes a spatiotemporal convolutional network, a bidirectional long short-term memory neural network, and a multi-head attention mechanism. The spatiotemporal convolutional network serves as a front-end feature extractor, the bidirectional long short-term memory neural network serves as a temporal dependency modeling module, and the multi-head attention mechanism serves as a global information integration module.

[0069] Step 3: Optimize the model parameters of the initial prediction model using the Magnificent Wren-Wren optimization algorithm. The learning rate, number of neurons in the spatiotemporal convolutional network, number of neurons in the bidirectional long short-term memory neural network, number of keys in the multi-head attention mechanism, and regularization parameters of the initial prediction model are used as optimization variables. The root mean square error of the prediction results of the initial prediction model on the dataset is used as the optimization objective function. The goal is to minimize the root mean square error to obtain the optimized prediction model.

[0070] Step 4, train the optimized prediction model: Divide the dataset into a training set and a test set, train the optimized prediction model until the model converges, and obtain the trained prediction model.

[0071] Step 5: Predict the remaining service life of the gas turbine using the post-trained prediction model: Collect the real-time operating parameters of the gas turbine to be predicted, extract the features of the operating parameters, process the extracted features, input the processed features into the post-trained prediction model, and the post-trained prediction model outputs the prediction result of the remaining service life of the gas turbine.

[0072] The parameters collected in steps 1 and 5 include low-pressure rotor speed, high-pressure rotor speed, power turbine speed, low-pressure compressor outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, low-pressure compressor outlet pressure, high-pressure compressor outlet pressure, low-pressure turbine outlet pressure, rated power, efficiency, and fuel type. The steps in steps 1 and 5 to extract features of the operating parameters include: extracting time-domain and frequency-domain features of the operating parameters. Time-domain features include maximum, minimum, average, peak, peak-to-peak, rectified average, variance, kurtosis, skewness, root mean square (RMS), waveform factor, peak factor, impulse factor, root square amplitude, margin factor, and kurtosis factor. Frequency-domain features include centroid frequency, MMS frequency, frequency variance, spectral entropy, and signal complexity.

[0073] The steps in steps 1 and 5 for processing the extracted features include performing sliding window sampling on the extracted features; the time span of the sliding window is 5-30 sampling periods, and the window sliding step size is 1 sampling period; the features extracted from the real-time running parameters are completely consistent with the feature types extracted from the whole life cycle data to ensure input distribution matching.

[0074] In step 4, during the training phase, the dataset is divided into a training set and a test set, with a ratio of 7:3 or 8:2. The optimized prediction model is trained using the Adam optimizer with a batch size of 32-128. The convergence criterion for the optimized prediction model is that the training loss decreases by less than 1×10⁻⁻⁶ over 50 consecutive iterations. 5 The total number of iterations reaches 500. Adam stands for Adaptive Moment Estimation.

[0075] In this embodiment, the extracted features are sampled using a sliding window and then denoised and normalized. In step 1, the denoised and normalized features are labeled with their corresponding remaining lifespan tags. In step 5, the denoised and normalized features are input into the trained prediction model.

[0076] The steps for running the trained prediction model in step 5 include:

[0077] Step 5.1: The spatiotemporal convolutional network uses causal convolutional layers, dilated convolutional layers, and residual connections to extract the memory history features of time-series data.

[0078] Step 5.2: The bidirectional long short-term memory neural network receives the memory history features of the time-series data, processes the positive and negative information of the memory history features of the time-series data, captures the global degradation dependency relationship of time, and fuses the features to output global information.

[0079] Step 5.3: The multi-head attention mechanism receives global information, assigns different importance weights to multiple independent attention heads, mines degradation time sequence information in the global information, and outputs the prediction result of the remaining service life of the gas turbine.

[0080] like Figure 2 As shown, the spatiotemporal convolutional network in this application includes multiple residual blocks, each residual block including a causal convolutional layer, a dilated convolutional layer, and residual connections; the causal convolutional layer is used to perform causal convolution operations on the temporal vector, the kernel size of the causal convolutional layer is 3×3, and the stride is 1; the dilated convolutional layer is used to perform dilated convolution operations on the output of the causal convolutional layer, and the dilation coefficient of the dilated convolutional layer is 2. a-1 The settings are configured such that 'a' is the convolutional layer number, a=1,2,3,4,5, to make the receptive field expand exponentially with the number of layers; the residual connection uses a shortcut connection method to fuse the output of each dilated convolutional layer with the input of the residual block, and uses an element-wise addition mode to integrate the input features and the convolutional output features; if the dimensions of the input and output are inconsistent, the dimensions are adjusted by 1×1 convolution before the addition operation is performed.

[0081] In one embodiment, the spatiotemporal convolutional network of the initial prediction model includes multiple residual blocks. Each residual block contains a main path and a shortcut path. The main path consists of two consecutive dilated causal convolutional units (DCUs) and is fused with the shortcut path via residual connections. Each DCU in the main path includes a DCU layer, a weight normalization layer, a ReLU activation function, and a Dropout layer connected in sequence. The DCU layer integrates a causal convolutional layer and a dilated convolutional layer. The residual connections fuse the output of the dilated convolutional layer, after processing by the ReLU activation function and the Dropout layer, with the input of the residual block. Here, ReLU stands for Rectified Linear Unit, and Dropout stands for Dropout Regularization.

[0082] Causal convolutional layers ensure that when processing time-series data, the current output depends only on the current and historical inputs, without revealing future information.

[0083] Weight normalization layers normalize the weight parameters of convolutional layers to accelerate model training and improve training stability. The ReLU activation function introduces a non-linear transformation into the network, enhancing its feature representation ability. Dropout layers randomly discard a portion of neurons during training, serving as a regularization technique to effectively prevent overfitting.

[0084] The residual connection is implemented through a shortcut path. Specifically, the final output of the second dilated causal convolutional unit in the main path is fused with the initial input of the residual block. The fusion uses an element-wise addition pattern to integrate the original input features and the convolutional output features. If the dimensions of the initial input and the final output are inconsistent, the shortcut path first adjusts the dimensions of the input features through a 1×1 convolutional layer before performing the addition operation.

[0085] like Figure 3 As shown, the bidirectional long short-term memory neural network comprises three layers of bidirectional LSTM units. Each layer consists of two sub-modules: a forward LSTM and a backward LSTM. The number of neurons in each forward LSTM and backward LSTM layer is the same. This multi-layered stacked structure can progressively deepen the understanding of gas turbine time-series operating data and improve the model's ability to remember and express complex degradation patterns.

[0086] In the prediction process using the trained prediction model, the bidirectional long short-term memory neural network obtains refined historical features of the time-series data by sharing the sequence of historical features of the output time-series data and processing them in reverse order. These features are then fused to output global information. Specifically, the forward LSTM traverses the input sequence in the forward temporal direction, updating cell and hidden states through gating units to generate a forward hidden state sequence. This effectively captures the forward temporal dependencies of gas turbine operating parameters, such as the progressive trend of performance degradation from normal. The backward LSTM traverses the same input sequence in the reverse temporal direction, updating cell and hidden states through gating units to generate a backward hidden state sequence, capturing the inverse temporal dependencies of the sequence. This process captures the inverse temporal dependencies of the time-series data, such as the retrospective impact of historical abnormal operating states on subsequent performance. In a single-layer bidirectional LSTM, the forward hidden state sequence generated by the forward LSTM and the backward hidden state sequence generated by the backward LSTM are fused. The fusion method typically involves performing element-weighted summation at corresponding time steps to obtain refined temporal features that include bidirectional contextual semantics. The refined features output from each layer of the bidirectional LSTM serve as input to the next layer. After three layers of this processing, the refined features from all time steps are concatenated to output global degradation feature information, providing a foundation for subsequent multi-head attention mechanisms to mine key degradation temporal information. LSTM stands for Long Short-Term Memory.

[0087] The bidirectional long short-term memory neural network of this application, through bidirectional processing of forward and backward LSTM and a multi-layer stacked structure, effectively alleviates the gradient vanishing problem of traditional recurrent neural networks and accelerates the convergence speed of training. On the other hand, it can more comprehensively and deeply capture the forward and backward time dependencies in the gas turbine operation sequence data, accurately mine the global degradation patterns contained in the data, and provide more reliable time-series feature support for the prediction of remaining service life.

[0088] The advantages of spatiotemporal convolutional networks lie in their high parallel computation efficiency, ability to capture long-distance dependencies, and avoidance of gradient problems inherent in long short-term memory (LSM) networks. Meanwhile, bidirectional LSM neural networks excel at handling the sequential dependencies of time-series data, making them suitable for analyzing the dynamic processes of performance degradation. This application combines the two, enabling them to extract local features from sensor data, such as transient anomalies, while also modeling long-term performance degradation trends, thereby improving the accuracy of gas turbine life prediction.

[0089] Traditional attention mechanisms may fail to fully utilize temporal relationships, resulting in poor model performance. Therefore, this application introduces a multi-head attention mechanism into the initial prediction model to extract time-series information, thereby improving the model's generalization ability.

[0090] The specific process of the multi-head attention mechanism in the prediction process using the trained prediction model is as follows:

[0091] Target Input Reception: The multi-head attention mechanism receives the global information from the feature fusion output, which serves as the target input for the multi-head attention mechanism. Feature Encoding: If it is the first layer of the multi-head attention mechanism, linear encoding is performed on the target input to generate a feature matrix H. If it is not the first layer, the output of the previous layer's multi-head attention is directly used as the feature matrix H. Multi-head Weighting: h independent linear projection matrices are constructed, where h ≥ 8. The linear projection matrices include a query projection matrix group Wq, a key projection matrix group Wk, and a value projection matrix group Wv. The feature matrix H is multiplied by each group Wq, Wk, and Wv respectively to obtain h query matrices Q1-Q. h H group bond matrix K1-K h h group value matrix V1-V h Single-head attention calculation: For each combination matrix (Q) m ,K m V m The attention weights are calculated using the scaled dot product attention formula, resulting in h groups of single-head attention outputs Z1-Z. h Where m takes the value of an integer from 1 to h; Multi-head result concatenation and projection: h groups of single-head attention outputs Z1-Z h The concatenation matrix Z_concat is obtained by concatenating the concatenation matrix along the column direction. Linear projection is then performed on the concatenation matrix Z_concat, which is multiplied by the learning weight matrix to reduce the dimensionality to the preset output dimension, resulting in the output matrix Z_final. Prediction results are output based on the projected output matrix Z_final, extracting global key degradation time series information and combining it with the preset remaining life mapping rules to output the prediction results of the remaining life of the gas turbine.

[0092] This application incorporates a multi-head attention mechanism, perfectly complementing spatiotemporal convolutional networks and bidirectional long short-term memory (LSSM) neural networks, constructing a classic model that leverages the strengths of both and avoids their weaknesses. The original spatiotemporal convolutional network has limitations: while it possesses a large receptive field through dilated convolutions, its ability to capture dependencies heavily depends on the kernel size and dilation coefficient, making it inefficient for very long and complex dependencies. Furthermore, the original bidirectional LSSM network has limitations: it exhibits sequential dependencies, and its computation is sequential, requiring step-by-step processing of the sequence. The current computation depends on the hidden state of the previous time step, limiting its parallel computing capabilities, especially when handling extremely long sequences, resulting in slow training speeds and the potential for gradient explosion in practice. The multi-head attention mechanism, however, offers the advantage of one-step global modeling. Matrix operations in the multi-head attention mechanism can be fully parallelized, independent of the sequential computation of the sequence. This allows the model to fully utilize the massively parallel computing capabilities of GPUs and other hardware during training, significantly accelerating the training speed. Any element in the sequence can be directly computed to be associated with any other element, regardless of the distance between them. This gives it an inherent advantage in capturing long-range dependencies, completely solving the inherent bottleneck of spatiotemporal convolutional network models.

[0093] This application utilizes a spatiotemporal convolutional network as its front end, which efficiently captures local features and short-term patterns, serving as an excellent feature extractor. A bidirectional long short-term memory neural network effectively learns the bidirectional contextual semantic information of sequences. Building upon the above processing, a multi-head attention mechanism enables global, focused information integration and decision-making. By considering all information provided by the front-end and middle-end layers—the outputs of the spatiotemporal convolutional network and the bidirectional long short-term memory neural network—the system dynamically determines which information is most critical to the current task and assigns different importance weights, ultimately making accurate predictions.

[0094] The optimization algorithm steps for the magnificent swan-warbler in step 3 of this application include:

[0095] Step 31: Convert the natural morphology of the magnificent swan warbler in nature into a mathematical model. Each individual magnificent swan warbler corresponds to a set of model parameters to be optimized in the initial prediction model. Each dimension corresponds to an optimization variable. Use a sinusoidal squared chaotic mapping to generate an initial candidate solution population in the search space. }, ;

[0096] The formula for the sinusoidal squared chaotic mapping is: ;

[0097] in, For the first The chaotic variable after the next iteration is initially set to a random decimal between 0 and 1. For the first The updated value after the next iteration; For control parameters; For the initial candidate solution population, the first The first individual Parameter values ​​for each dimension; This is the upper bound of the search space; This is the lower bound of the search space.

[0098] This application reconstructs the initialization process by introducing a sinusoidal square chaotic mapping. By utilizing the ergodicity and randomness of the chaotic mapping, a more uniform and higher-quality initial candidate solution population can be generated in the solution space, laying a solid foundation for the global optimization of the algorithm and improving optimization efficiency.

[0099] Step 32: Determine whether an individual belongs to a juvenile bird based on the optimization objective function value corresponding to each individual in the initial candidate solution population. That is, if the root mean square error of the prediction result of the initial prediction model corresponding to each model parameter combination in the initial candidate solution population is greater than a preset value, then the individual corresponding to the current model parameter combination is determined to be a juvenile bird. Determine the proportion of juvenile birds in the initial candidate solution population. If the proportion of juvenile birds is greater than 0.5, then update the individual parameter values ​​in the initial candidate solution population by simulating the growth stage of juvenile magnate warblers until the proportion of juvenile birds in the updated initial candidate solution population is less than or equal to 0.5, thus obtaining a qualified candidate solution population.

[0100] The updated formula is: ,in, The first candidate solution in the initial candidate solution population After the (t+1)th iteration, the individual... The updated parameter values ​​for each dimension. Indicates the first The parameter values ​​in the j dimensions of an individual after t iterations, where rand is a random number between [0,1]. This is the upper limit of the decision variables in the initial candidate solution population. This serves as the lower bound for decision variables in the initial candidate solution population. The process simulates how fledglings update their positions by learning from extensive growth experiences. The aim is to significantly improve the algorithm's exploration capability in the early stages of the global search by introducing a large-scale random positional perturbation, enabling members of the Magnificent Wren to make extensive positional changes.

[0101] If the proportion of juvenile birds in the initial candidate solution population is less than or equal to 0.5, then the initial candidate solution population is directly regarded as a qualified candidate solution population.

[0102] Step 33: Calculate the risk threshold s in the qualified candidate solution population. If the risk threshold is less than a preset value, proceed to step 34; if the risk threshold is greater than or equal to the preset value, proceed to step 35. , and It is a random number with a normal distribution, and the default value in this application is 20.

[0103] Step 34: By simulating the teaching mechanism of the Magnificent Swan-Wren during breeding and rearing, the optimal solution individual is selected from the qualified candidate solution population. The model parameter combination corresponding to the optimal solution individual is used as the model parameters of the optimized prediction model. Specifically, the individual with the optimal objective function in the qualified candidate solution population is taken as the subject, that is, the individual with the smallest root mean square error of the prediction result of the initial prediction model corresponding to the model parameter combination in the qualified candidate solution population on the dataset is taken as the subject. The parameters of the remaining individuals in the qualified candidate solution population are adjusted according to the position update rules of simulating the breeding and rearing process of the Magnificent Swan-Wren, and the parameter value combination of each individual is updated. The update formula is: ;

[0104] in, The first qualified candidate solution in the population The first individual The updated parameter values ​​for each dimension. This refers to the optimal combination of individual parameter values ​​for optimizing the objective function within the qualified candidate solution population. express Individuals t After the second iteration j The parameter values ​​of each dimension, The constant value is 0.8;

[0105] It is defined as a factor that matures as the teaching cycle shortens. , ;

[0106] in, The expression for the loop, This is the upper limit of the decision variables in the qualified candidate solutions. This represents the lower bound of the decision variables among the qualified candidate solutions. For the current number of assessments, This represents the maximum number of evaluations.

[0107] If the objective function corresponding to the updated parameter value combination of the remaining individuals in the qualified candidate solution population is better than the objective function corresponding to the original parameter value combination, that is, the root mean square error of the prediction result of the initial prediction model corresponding to the updated parameter value combination on the dataset is smaller, then the updated parameter value combination of the individual replaces the original parameter value combination of the individual. After all the remaining individuals complete the preset number of updates, a new qualified candidate solution population is obtained. From the new qualified candidate solution population, the individual with the optimal objective function is selected, and the model parameter combination corresponding to the optimal individual is used as the model parameters of the final optimized prediction model.

[0108] In this process, member positions are updated by simulating the cooperative teaching mechanism of the Magnificent Slender-tailed Warbler during breeding and rearing. This behavior is modeled as a small positional adjustment near the current optimal solution. This refined movement strategy increases the algorithm's local search capability within known high-quality solution regions, contributing to precise convergence towards the optimal solution.

[0109] Step 35: Select the optimal solution individual from the qualified candidate solution population by simulating the mechanism of the magnificent slender-tailed warbler in avoiding predators. The model parameter combination corresponding to the optimal solution individual is used as the model parameters of the optimized prediction model.

[0110] Step 35 specifically includes:

[0111] The individual with the optimal objective function in the qualified candidate solution population is taken as the subject, that is, the individual with the smallest root mean square error of the prediction result of the initial prediction model corresponding to the model parameter combination in the qualified candidate solution population is taken as the subject. The parameters of the remaining individuals in the qualified candidate solution population are adjusted according to the position update rules of the simulated magnificent swan warbler in the process of avoiding predators, and the parameter value combination of each individual is updated.

[0112] The updated formula is: ;

[0113] in, The first qualified candidate solution in the population The first individual The updated parameter values ​​for each dimension. This refers to the optimal combination of individual parameter values ​​for optimizing the objective function within the qualified candidate solution population. express Individuals t After the second iteration j Parameter values ​​for each dimension;

[0114] The random step size during flight. , for The random values ​​in the input are introduced into the spiral search factor. Dynamic adjustments are made to make the random step size during flight exhibit a spiral expansion characteristic in the solution space;

[0115] For adaptive flight balance factor, , , For call frequency values, Indicates the current evaluation number. Indicates the maximum number of evaluations.

[0116] If the objective function corresponding to the updated parameter value combination of the remaining individuals in the qualified candidate solution population is better than the objective function corresponding to the original parameter value combination, that is, the root mean square error of the prediction result of the initial prediction model corresponding to the updated parameter value combination on the dataset is smaller, then the updated parameter value combination of the individual replaces the original parameter value combination of the individual. After all the remaining individuals complete the preset number of updates, a new qualified candidate solution population is obtained. From the new qualified candidate solution population, the individual with the optimal objective function is selected, and the model parameter combination corresponding to the optimal individual is used as the model parameters of the final optimized prediction model.

[0117] This application simulates the behavior of the Magnificent Slender-tailed Warbler in avoiding predators, updating the population position based on two different defense mechanisms the warbler employs when attacked by predators. Individuals directly attacked quickly escape, resulting in small, fine-tuned adjustments to their positions; while other members circle and remain vigilant, causing large-scale changes in their positions. These two differentiated movement patterns enhance the algorithm's local exploitation capability and global exploration range, respectively. The random step size during flight directly affects the exploration and exploitation capabilities at this stage. To overcome the tendency to get trapped in local optima, a spiral search factor is introduced. Dynamic adjustments are made to induce a spiral expansion characteristic within the solution space, thereby achieving an adaptive balance between global exploration and local development. It adds control over the direction of individual movement to prevent individuals from moving towards unfavorable positions and wasting assessment time.

[0118] In this embodiment of the application, the value range of the optimization variables is: learning rate ∈ [0.0001, 0.01], number of neurons in spatiotemporal convolutional network ∈ [32, 256], number of neurons in bidirectional long short-term memory neural network ∈ [32, 256], number of multi-head attention keys ∈ [16, 128], and regularization parameter ∈ [0.0001, 0.01].

[0119] This application introduces a sinusoidal squared chaos model for population initialization into the optimization algorithm for the magnificent swan-warbler, thereby addressing the issue of low population diversity and ensuring a more uniform distribution in the search space, leading to solutions with higher applicability. Simultaneously, a spiral search mechanism is employed to improve the optimizer, forming a periodic circular motion in the solution space from far to near and from near to far, dynamically adjusting the search step size according to the actual situation, thus balancing global exploration and local exploitation. Furthermore, it can switch between spiral and linear updates with a certain probability, effectively reducing the risk of premature convergence and further improving robustness. Integrating the improved optimization algorithm into the model enhances global exploration in the early stages and strengthens local exploitation in the later stages, thereby finding better solutions in a shorter time and avoiding getting trapped in local optima. Additionally, it reduces the model's sensitivity to initial parameters. Since the original algorithm is quite sensitive to the initial candidate solution population distribution and control parameters, the improved algorithm maintains stable performance under different problem scenarios through an adaptive parameter adjustment mechanism. Ultimately, by adopting the improved Magnificent Slender-tailed Warbler optimization algorithm, it is possible to accelerate model training and improve the prediction accuracy of gas turbines.

[0120] Experimental verification shows that the lifetime prediction curve of the proposed method decreases synchronously with the actual remaining lifetime without directional deviation, and the relative error is within 5%. Compared with mainstream models, its mean absolute error is reduced by 1.9% and 11.5% compared with short-term memory neural networks and gated recurrent unit algorithms, respectively, and its mean squared error is the smallest, demonstrating that the proposed method has excellent prediction accuracy in both long- and short-cycle degradation scenarios of gas turbines. Furthermore, the proposed method reduces the training time per cycle by 39% compared with the gated recurrent unit algorithm, further verifying its significant efficiency advantage.

[0121] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the remaining service life of a gas turbine based on an improved SFOA algorithm, characterized in that, include: Step 1: Collect the operating parameters of the gas turbine throughout its entire life cycle, extract the features of the operating parameters, process the extracted features, and label the processed features with the corresponding remaining service life labels to form a dataset; Step 2: Construct an initial prediction model, which includes a spatiotemporal convolutional network, a bidirectional long short-term memory neural network, and a multi-head attention mechanism. The spatiotemporal convolutional network serves as a front-end feature extractor, the bidirectional long short-term memory neural network serves as a temporal dependency modeling module, and the multi-head attention mechanism serves as a global information integration module. Step 3: Optimize the model parameters of the initial prediction model using the Magnificent Wren-Wren optimization algorithm. The learning rate, number of neurons in the spatiotemporal convolutional network, number of neurons in the bidirectional long short-term memory neural network, number of keys in the multi-head attention mechanism, and regularization parameters of the initial prediction model are used as optimization variables. The root mean square error of the prediction results of the initial prediction model on the dataset is used as the optimization objective function. The goal is to minimize the root mean square error to obtain the optimized prediction model. Step 4: Divide the dataset into training and testing sets, train the optimized prediction model until the model converges, and obtain the trained prediction model. Step 5: Collect the real-time operating parameters of the gas turbine to be predicted, extract the features of the operating parameters, process the extracted features, input the processed features into the trained prediction model, and the trained prediction model outputs the prediction result of the remaining service life of the gas turbine. The steps for running the trained prediction model in step 5 include: Step 5.1: The spatiotemporal convolutional network uses causal convolutional layers, dilated convolutional layers, and residual connections to extract the memory history features of time-series data; Step 5.2: The bidirectional long short-term memory neural network receives the memory history features of time-series data, processes the positive and negative information of the memory history features of time-series data, captures the global degradation dependency relationship of time, and fuses the features to output global information. Step 5.3: The multi-head attention mechanism receives global information, assigns different importance weights to multiple independent attention heads, mines degradation time sequence information in the global information, and outputs the prediction result of the remaining service life of the gas turbine. The optimization algorithm steps for the magnificent slender-tailed warbler in step 3 include: Step 31: Convert the natural morphology of the magnificent swan warbler in nature into a mathematical model. Each individual magnificent swan warbler corresponds to a set of model parameters to be optimized in the initial prediction model. Each dimension corresponds to an optimization variable. Use a sinusoidal squared chaotic mapping to generate an initial candidate solution population in the search space. }, ; The formula for the sinusoidal squared chaotic mapping is: ; in, For the first The chaotic variable after the next iteration is initially set to a random decimal between 0 and 1. For the first The updated value after the next iteration; For control parameters; For the initial candidate solution population, the first The first individual Parameter values ​​for each dimension; This is the upper bound of the search space; This serves as the lower bound of the search space. Step 32: Determine the proportion of juveniles in the initial candidate solution population. If the proportion of juveniles is greater than 0.5, update the individual parameter values ​​in the initial candidate solution population by simulating the growth stage of juvenile magnate warbler until the proportion of juveniles in the updated initial candidate solution population is less than or equal to 0.5, and obtain a qualified candidate solution population. If the proportion of juveniles in the initial candidate solution population is less than or equal to 0.5, then the initial candidate solution population is directly regarded as a qualified candidate solution population. Step 33: Calculate the risk threshold s in the qualified candidate solution population. If the risk threshold is less than a preset value, proceed to step 34; if the risk threshold is greater than or equal to the preset value, proceed to step 35. , and It is a random number that follows a normal distribution; Step 34: Select the optimal solution individual from the qualified candidate solution population by simulating the teaching mechanism of the magnificent slender-tailed warbler in the process of breeding and feeding. The model parameter combination corresponding to the optimal solution individual is used as the model parameters of the optimized prediction model. Step 35: Select the optimal solution individual from the qualified candidate solution population by simulating the mechanism of the magnificent slender-tailed warbler in avoiding predators. The model parameter combination corresponding to the optimal solution individual is used as the model parameters of the optimized prediction model.

2. The gas turbine remaining service life prediction method based on the improved SFOA algorithm according to claim 1, characterized in that, The spatiotemporal convolutional network includes multiple residual blocks, each residual block including a causal convolutional layer, a dilated convolutional layer, and residual connections; The causal convolutional layer is used to perform causal convolution operations on the temporal vector. The kernel size of the causal convolutional layer is 3×3 and the stride is 1. The dilated convolutional layer is used to perform dilated convolution operations on the output of the causal convolutional layer, and the dilation coefficient of the dilated convolutional layer is set at 2. a-1 The settings are configured such that a is the convolutional layer number, a=1,2,3,4,5, so that the receptive field expands exponentially with the number of layers; The residual connection uses a shortcut connection method to fuse the output of each dilated convolutional layer with the input of the residual block, and integrates the input features and convolutional output features using an element-wise addition mode; if the dimensions of the input and output are inconsistent, the dimensions are adjusted first by a 1×1 convolution before the addition operation is performed.

3. The method for predicting the remaining service life of a gas turbine based on the improved SFOA algorithm according to claim 1, characterized in that, In step 5.2, the bidirectional long short-term memory neural network obtains the refined memory history features of the time-series data by sharing the sequence of memory history features of the output time-series data and processing them in reverse time order. Finally, the features are fused and the global information is output. The bidirectional long short-term memory neural network contains three layers of bidirectional LSTM units, with the same number of neurons in each forward LSTM and backward LSTM layer. The forward LSTM traverses the input sequence in the forward temporal direction, updates the cell state and hidden state through the gating unit, generates a forward hidden state sequence, and captures the forward temporal dependence of the sequence. The backward LSTM traverses the same input sequence in the reverse temporal direction, updates the cell state and hidden state through the gating unit, generates the reverse hidden state sequence, and captures the reverse temporal dependence of the sequence.

4. The gas turbine remaining service life prediction method based on the improved SFOA algorithm according to claim 1, characterized in that, Step 5.3 specifically includes: Target input reception: The multi-head attention mechanism receives the global information output by feature fusion and uses it as the target input for the multi-head attention mechanism; Feature encoding: If it is the first layer of multi-head attention mechanism, linear encoding is performed on the target input to generate feature matrix H; if it is not the first layer of processing, the output of the previous layer of multi-head attention is directly used as feature matrix H. Multi-head weighting: Construct h independent linear projection matrices, h≥8. The linear projection matrices include query projection matrix group Wq, key projection matrix group Wk, and value projection matrix group Wv. Multiply the feature matrix H by each group Wq, Wk, and Wv respectively to obtain h query matrices Q1-Q. h H group bond matrix K1-K h h group value matrix V1-V h ; Single-head attention calculation: For each combination matrix (Q) m ,K m V m The attention weights are calculated using the scaled dot product attention formula, resulting in h groups of single-head attention outputs Z1-Z. h , where m takes the value of an integer from 1 to h; Multi-head result stitching and projection: The h groups of single-head attention outputs Z1-Z h Concatenate the columns to obtain the concatenation matrix Z_concat; perform linear projection on the concatenation matrix Z_concat, multiply the concatenation matrix Z_concat by the learned weight matrix, reduce the dimension to the preset output dimension, and obtain the output matrix Z_final; Prediction results output: Based on the projected output matrix Z_final, extract global key degradation time series information, and combine it with the preset remaining life mapping rules to output the prediction results of the remaining life of the gas turbine.

5. The method for predicting the remaining service life of a gas turbine based on the improved SFOA algorithm according to claim 1, characterized in that, Step 35 specifically includes: Taking the individual with the best optimization objective function in the qualified candidate solution population as the subject, the parameters of the remaining individuals in the qualified candidate solution population are adjusted according to the position update rules of the magnificent swan warbler in the process of simulating the mechanism of avoiding natural enemies, and the parameter value combination of each individual is updated. The updated formula is: ; in, The first qualified candidate solution in the population The first individual The updated parameter values ​​for each dimension. This refers to the optimal combination of individual parameter values ​​for optimizing the objective function within the qualified candidate solution population. express Individuals t After the second iteration j Parameter values ​​for each dimension; The random step size during flight. , for Random values ​​in; For adaptive flight balance factor, , , For call frequency values, Indicates the current evaluation number. Indicates the maximum number of evaluations; If the objective function corresponding to the updated parameter value combination of the remaining individuals in the qualified candidate solution population is better than the objective function corresponding to the original parameter value combination, then the updated parameter value combination of the individual replaces the original parameter value combination of the individual. After all the remaining individuals complete the preset number of updates, a new qualified candidate solution population is obtained. From the new qualified candidate solution population, the individual with the optimal objective function is selected, and the model parameter combination corresponding to the optimal individual is used as the model parameters of the final optimized prediction model.

6. The method for predicting the remaining service life of a gas turbine based on the improved SFOA algorithm according to claim 1, characterized in that, The steps in steps 1 and 5, which involve extracting features of the operating parameters, include: Extract the time-domain and frequency-domain features of the operating parameters. The time-domain features include maximum value, minimum value, average value, peak value, peak-to-peak value, rectified average value, variance, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, root mean square amplitude, margin factor, and kurtosis factor. Frequency domain characteristics include centroid frequency, mean square frequency, frequency variance, spectral entropy, and signal complexity.

7. The method for predicting the remaining service life of a gas turbine based on the improved SFOA algorithm according to claim 1, characterized in that, In steps 1 and 5, the step of processing the extracted features includes performing sliding window sampling on the extracted features; The sliding window spans 5-30 sampling periods, and the window sliding step is 1 sampling period. The features extracted from real-time operating parameters are completely consistent with the feature types extracted from the whole life cycle data, ensuring that the input distribution matches.

Images